AUTOMATIC DSM GENERATION FROM LINEAR ARRAY IMAGERY DATA

Zhang Li, Armin Gruen

Institute of Geodesy and Photogrammetry, Swiss Federal Institute of Technology ZurichETH−Hoenggerberg; CH-8093 Zurich, Switzerland

Tel.: +41-1-633 31 57, Fax: +41-1-633 11 01E-mail: <zhangl><agruen>@geod.baug.ethz.ch

Commission III, WG III/2

KEY WORDS: Linear Array Imagery, Image Matching, DSM

ABSTRACT:CCD linear array sensors are widely used to acquire panchromatic and multispectral imagery for photogrammetric and remotesensing applications. The processing of this kind of images provides a challenge for algorithmic redesign and this opens thepossibility to reconsider and improve many photogrammetric processing components. In addition, the basic capabilities of imagematching techniques have so far not been fully utilized yet. This paper presents a matching procedure for automatic DSM generationfrom linear array imagery data. It can provide dense, precise and reliable results. The method uses a coarse-to-fine hierarchicalsolution with an effective combination of several image matching algorithms and automatic quality control. The DSMs are generatedby combination of matching results of feature points, grid points and edges. Finally, a modified multi-photo geometricallyconstrained (MPGC) matching algorithm is employed to achieve sub-pixel accuracy for all the matched features with multi-image ormulti-strip image data.The proposed approach in this paper has been applied to different areas with varying textures and terrain types. The accuracy tests arebased on the comparison between the high quality DEMs / DSMs derived from airborne Laser Scanner or manual measurements andthe automatic extracted DSMs. Results with STARIMAGER, IKONOS and SPOT5 HRS images are reported. We demonstrate withthese experiments that our approach leads to good results.

1. INTRODUCTIONIn recent years, CCD linear array sensors are widely used toacquire panchromatic and multispectral imagery in pushbroommode for photogrammetric and remote sensing applications.Linear scanners are carried on aircraft (e.g. ADS40), helicopter(e.g. STARIMAGER) or spacecraft (e.g. IKONOS) and allowfor photogrammetric mapping at different scales.Spaceborne optical sensors like SPOT, IKONOS, and QuickBirdprovide not only for high-resolution (0.6 – 5.0 m) and multi-spectral data, but also for the capability of stereo mapping. Therelated sensors are all using linear array CCD technology forimage acquisition and are equipped with high quality orbitposition and attitude determination devices like GPS and IMUsystems.Progress in the development of airborne linear array imagingsystem has also been made in the last decade. These systems usethe three-line-scanner concept and provide for high resolution(0.5 – 0.03 m) panchromatic and multispectral image data withtriplet overlap and along-track base direction. In the year 2000,Starlabo Corporation, Tokyo designed a new airborne digitalimaging system, the Three-Line-Scanner (TLS) system (nowcalled STARIMAGER (SI)), jointly with the Institute ofIndustrial Science, University of Tokyo (Murai, Matsumoto,2000). The first generation camera STARIMAGER-100 (SI-100) contains three parallel one-dimensional CCD focal planearrays, with 10200 pixels of 7µm each. Starlabo is currentlydeveloping a new generation camera system SI-200. This comeswith an improved lens system and with 10 CCD arrays on thefocal plane (3 × 3 work in RGB mode, 1 CCD array works ininfrared mode). Each CCD array consists of 14 404 pixels at5µm size. The system produces seamless high-resolution images(3 - 10 cm footprint on the ground) with three viewing directions(forward, nadir and backward). For the SI sensor and imagingparameters see Gruen, Zhang, 2002.The processing of this kind of images provides a challenge foralgorithmic redesign and this opens the possibility to reconsiderand improve many photogrammetric processing components,like image enhancement, multi-channel color processing,triangulation, orthophoto and DEM generation and objectextraction. We have recently developed a full suite of newalgorithms and software system for the precision processing ofthis kind of data.In this paper, we put particular emphasis on the automaticgeneration of DSMs. Originally we developed a matchingapproach and the related software “SI-Matcher” for multi-image

processing of the very high-resolution SI images (Gruen, Zhang,2003). Now this matching procedure has been extended and hasthe ability to process other linear array images as well. We willbriefly report about the basic considerations for our procedure.Then we will address the key algorithms. We will giveexperimental results from the processing of SI, IKONOS andSPOT5 HRS images.

2. MATCHING CONSIDERATIONSAutomatic DEM/DSM generation through image matching hasgained much attention in the past years. A wide variety ofapproaches have been developed, and automatic DEMgeneration packages are in the meanwhile commerciallyavailable on several digital photogrammetric workstations.Although the algorithms and the matching strategies used maydiffer from each other, the accuracy performance and theproblems encountered are very similar in the major systems andthe performance of commercial image matchers does by far notlive up to the standards set by manual measurements (Gruen etal., 2000). The main problems in DEM/DSM generation areencountered with

(a) Little or no texture(b) Distinct object discontinuities(c) Local object patch is no planar face(d) Repetitive objects(e) Occlusions(f) Moving objects, incl. shadows(g) Multi-layered and transparent objects(h) Radiometric artifacts like specular reflections and others(i) Reduction from DSM to DEM

The degree to which these problems will influence the matchingresults is imagescale-dependent. A DSM derived from 5 mpixelsize SPOT5 HRS images or 1 m pixelsize IKONOS imageswill be relatively better than one derived from 5 cm pixelsize SIimages. To extract DSMs from very high-resolution aerialimages, we should take into account the occlusions, the surfacediscontinuities such as man-made objects and trees, large areaswith little or even no texture, repetitive patterns, etc.On the other hand, linear array imagery provides for newcharacteristics and possibilities for image matching:● It has the ability to provide 16 bit (effectively 12 bit) images,which should reduce the number of mismatches even in darkshadow areas.

● It has the ability to provide multiple images with multiplechannels. Thus it enables the multi-image matching approach,which leads to a reduction of problems caused by occlusions,multiple solutions, surface discontinuities and results in highermeasurement accuracy through the intersection of more than twoimage rays.● It has the ability to provide for relatively precise orientationelements that can be used to enforce geometric constraints andrestrict the search space along quasi-epipolar lines.●The nearly parallel projection in along-track direction causesless occlusion on the nadir-viewing images.

Figure 1: Workflow of our image matching procedure

Among the known matching techniques and algorithms, area-based (ABM) and feature-based matching (FBM) are the twomain ones applied to automatic DSM generation in general.ABM and FBM have both advantages and disadvantages withrespect to the problems presented above. The key to successfulmatching is an appropriate matching strategy, making use of allavailable and explicit knowledge concerning sensor model,network structure and image content.Our matching approach is a hybrid method that combines ABMand FBM. It aims to generate DSMs by attacking the problems(a)-(f) mentioned above. It uses a coarse-to-fine hierarchicalsolution with a combination of several image matchingalgorithms and automatic quality control. Figure 1 shows theworkflow of our matching procedure. After the image pre-processing and production of the image pyramid, the matches ofthree kinds of features, i.e. feature points, grid points and edgeson the original images are finally found progressively startingfrom the low-density features on the images with low resolution.A triangular irregular network (TIN) based DSM is constructedfrom the matched features on each level of the pyramid, which inturn is used in the subsequent pyramid level for theapproximations and adaptive computation of the matchingparameters. Finally the modified MPGC matching is used toachieve more precise matches for all the matched features andidentify some false matches. In the MPGC procedure, multiplestrip image data can be introduced and combined. More detailsof our approach will be provided in the next paragraph.

3. THE MATCHING APPROACH3.1 Image PreprocessingIn order to reduce the effects of radiometric problems like strongbright and dark regions and to optimise the images forsubsequent feature extraction and image matching, a pre-processing method, which combines an edge-preservingsmoothing filter and the Wallis filter. First, the edge preservingsmoothing filter proposed by Saint-Marc et al., 1991 was appliedto reduce the noise, while sharpening edges and preserving evenfine detail such as corners and line endpoints. Next, the Wallisfilter, which strongly enhances already existing texture patterns,is applied. The examples of Figure 2 indicates that even in

shadow and “homogeneous” areas the image content isenhanced.The image pyramid starts from the original resolution images.Each pyramid level is generated by multiplying a generationkernel and reduces the resolution by factor 3. The pyramid levelnumber is a pre-defined value that is either a user-input or can bedetermined according to the height range of the imaging area.

Figure 2: SI image before and after pre-processing

3.2 Feature Point MatchingWe use the Foerstner interest operator to extract well-definedfeature points that are suitable for image matching. Firstly thereference image is divided into small image patches (the nadirviewing SI or satellite image is selected as reference). Only onefeature point will be extracted in each image patch. The densityof the feature points can be controlled by the size of the patches.The determination of the correspondences of the given points onthe search images is carried out using the geometricallyconstrained cross-correlation method (see Gruen, Zhang, 2003).The matching candidates are computed by cross-correlation witha set of adaptively determined parameters like the image windowsize ws, the threshold of the correlation coefficient ct and thesearch distance. The approximate DSM that is derived from thehigher-level of the image pyramid is used to estimate theseparameters.We incorporate the method proposed by Kanade & Okutomi,1994 to select an optimal window size by evaluating the imagecontent and the disparity within the matching window. As aresult, in flat areas with small image intensity variations, thewindow size ws increases and in areas of large terrain elevationvariations it decreases. The threshold of the correlationcoefficient ct should also vary according to the roughness of theterrain. We set a larger value in flat areas and smaller value inrough terrain areas. The search window size depends on theterrain elevation variation in a small neighborhood of the givenpoint and on the image geometry. In relatively flat areas the sizeof the search window decreases and vice versa.By adaptive selection of these parameters, we can both reducethe processing time and the probability for multiple matchingcandidates. The number of matching candidates can be furtherreduced by introducing a third image. For every candidate, itsposition on the third image can be predicted using the imageorientation elements. If the correlation coefficient between thereference and the third image is lower than the threshold, thismatching candidate will be dropped. However, we cannotcompletely avoid the ambiguity problem due to reasons likerepetitive patterns. Our procedure takes n (≤ 5) matchingcandidates with correlation coefficient values above thethreshold ct.As a result, for each feature point on the reference image severalmatching candidates can be computed. The correct match isdetermined by analysing the following quality measuressequentially:a) The correct match should have a clear and sharp maximalcorrelation coefficient. If there are more than one candidates andthe value of the first correlation coefficient peak is more than twotimes of that of the secondary peak, the candidate that has thelargest correlation coefficient value will be considered the correctmatch.b) Using the same matching parameters, the feature point can beback-matched from the search images to the reference image. Ifthe difference between this two-way matching is less than onepixel, the candidate is assumed to be the correct match.c) Under the condition that the feature point appears on morethan two images, the residuals of the photogrammetric forwardintersection should be less than 2-3 times the standard deviationsof image coordinates of the triangulation adjustment.

The points that pass these tests will be indicated as reliablematches. The points that cannot pass these tests may havemultiple solutions. This matching ambiguity will be solved bythe following global image matching method through localsmoothness constraints (see chapter 3.4).

3.3 Edge MatchingTo reconstruct a DSM from very high-resolution images overurban areas we must take into account the problems caused bysurface discontinuities, occlusions and the significant perspectiveprojection distortion. Even with satellite images, line features arealso important for capturing and modeling terrain features suchas ridgelines and breaklines. Matching the edges is a possiblesolution to these problems. However we should consider thefollowing problems:• The edge on one image may break into more than one segmentdue to image noise, occlusions and the deficiencies of the featureextraction algorithms.• The conjugate edges on different images may have quitedifferent shapes due to the projection distortion.• There may be many similar features in a search area.The edge matching procedure presented here is based on theevaluation of the local geometric and photometric attributes ofedges for the solution of disambiguities. The quasi-epipolargeometry and the DSM data derived from the higher-level of theimage pyramid are used to provide for the matching candidatesfor each edgel. A figural continuity constraint satisfactionscheme (the disparity along an edge should change smoothly)and a shape matching approach are used to achieve the finalresults.The well-known Canny operator is used to locate the intensitydiscontinuities. Then edgels are linked into free-form edgesthrough a local processing that analyses the characteristics ofthese pixels in a small neighborhood. This approach is carriedout independently on three images. Only edges above aminimum length (30 pixels for SI images and 15 pixels forsatellite images) are considered for matching.The edgels along the given edge are matched with the edges thatare defined at the intersection points between the candidateedges and the correspondend epipolar curve within the searchwindow on one of the search images. The search window can bedetermined by using the same method as in chapter 3.2. Theremay be several matching candidates within the search window.To solve this ambiguity problem we perform the following threesteps sequentially:a) Evaluation of the difference of the local edge orientation

between the given edgel and its candidates. The localorientation for an edgel is the image intensity gradientcomputed modulo 2π. Candidates with differences above athreshold will be dropped. Considering the possible reliefdistortion, this constraint should not be too tight. Forexample, we use 40 degrees for SI images.

b) Evaluation of the normalized cross-correlation coefficientof the intensity values on each side of the edgel. Exclusionof those candidates that have a very low correlationcoefficient, e.g. less than 0.5.

c) If more than two images are available, each candidate canbe validated on the third image through the indicator usedin step b).

After these three steps, the given edgel may still have more thanone candidate. The problem will be further solved by using thefigural continuity constraint through a relaxation method. Thismethod examines the candidates by computing how muchsupport they receive from their local neighborhood along theedge. We select the candidate that gains the highest support asthe correct match for each edgel. For each edge, the edgels thathave only one candidate will serve as “anchors” for thisrelaxation method. By linking the successfully matched edgels inthe search images, we obtain the correspondent edge matc.Finally, we do a shape matching between the given edge and itscorrespondent edge through least squares adjustment. Only theedges with small shape matching errors will be kept.Figure 3 shows an example of our edge matching.

Figure 3: Edge matching with SI images (5 cm resolution)

3.4 Grid Point MatchingWe use grid points to create uniformly distributed points overthe whole images even in areas with very little or no texture. Thecorrespondences of these grid points could be computed by usingthe method presented in chapter 3.2. Compared to the featurepoints, the choice of grid points is blind and thus many gridpoints are located in areas with weak or no texture. The searchfor the match of a given grid point has more possibilities to yieldmultiple candidates, or even no candidate.To solve this problem, we use a global image matching methodwith relaxation technique. This method examines the candidatesby computation how much support they receive from their localneighborhood and select the candidate that gains the highestsupport as the correct match. Here we use Prazdny’s “coherenceprinciple” model (Prazdny, 1985). We incorporate this idea inour global image matching and get the solution by relaxationtechnique.Firstly, the points are selected in form of a regular grid in thereference image. Their matching candidates on the search imagesare computed. Together with all the matched feature points andedges they construct a TIN. It should be noted here that all thematched points can be categorized into three classes: Pointshaving reliable matches, points having several candidates, andpoints without matching candidates. In the first case, they aretreated as having only one matching candidate and, they serve as“anchors” for the global matching procedure. For the last case,they will be given several “false” candidates (with a very smallcorrelation coefficient value) evenly distributed within the searchwindow. The matched edges serve as break-lines in the TINstructure. They control the weights of the local smoothnessconstraints.This method is performed on stereo pairs. The key point of thismethod, that distinguishes it from the single point matching, isits compatible coefficient function and its smoothness constraintsatisfaction scheme. With the smoothness constraint, areas withhomogeneous or only little texture can be bridged over,assuming that the terrain surface varies smoothly over the area.In the meantime, the surface discontinuities can be preservedbecause the smoothness constraints cannot cross the edges. Fordetails of this procedure see (Gruen, Zhang, 2003).

3.5 Matching Through the Image PyramidsA triangular irregular network (TIN) based DSM is constructedfrom the matched features on each level of the pyramid, which inturn is used in the subsequent pyramid level for theapproximations and adaptive computation of the matchingparameters. The matched edges are used as breaklines such thatno triangle crosses these edges. The TIN maintains the originalmatching results without any interpolation. The surfacediscontinuities of the terrain can be well captured and preserved.The initial DSM for the highest level of image pyramid can beextracted by standard cross-correlation based on a “regiongrowing” matching strategy. This method uses the alreadymeasured control and tie points as seed points and matches thepoints under the assumption that points in a local neighborhoodshould have similar disparities (Otto, Chau, 1988). This methodis justified because the disparity surface can be treated ascontinuous and smooth on the lowest resolution image pyramidlevel. In some difficult areas like very rough alpine terrain, some

manually measured points can also be introduced as seed points.It will give better approximations for the matching.

3.6 Refined Matching Based on the Modified MPGCMPGC (Multi-Photo Geometrically Constrained Matching) wasdeveloped by Gruen, 1985, and is described in detail inBaltsavias, 1991. It combines least squares matching andgeometric constraints formulated either in image or in objectspace and permits a simultaneous determination of pixel andobject coordinates. Any number of images (more than two) canbe used simultaneously. The achieved accuracy is in the sub-pixel range.Our modified algorithm is an extension of the standard MPGC. Itintegrates the geometric constraints derived from the LinearArray sensor models. The geometric constraints force thematching to search for a conjugate point only along a small bandaround the epipolar curve and reduce the possibility of falsematches.

Figure 4: MPGC matching with multi-strip SI images.Top: Images of a strip acquired from west to east

Bottom: Images of the cross-strip

The modified MPGC is used to refine the matched features inorder to achieve sub-pixel accuracy. The DSM derived from theapproaches (3.2)-(3.4) provides quite good approximations forthe MPGC procedure and increases the convergence rate. Theinitial values of the shaping parameters in MPGC can bepredetermined by using the image geometry and the derivedDSM data. The corresponding rays of the four corners of thematching window in the reference image are intersected with thederived DSM and their object coordinates can be determined.Through imaging geometry, the corresponding imagecoordinates in the search images can also be determined. Theinitial values of the shaping parameters can easily be computedfrom these four corner points and their correspondences. By thisway, features can be matched with multiple images or even withmultiple image strips that have different flight directions andimage scales. Figure 4 gives an example for the case of cross-strips.For edge matching the parameters of a spline function of the 3Dobject edge are directly estimated together with the matchingparameters of edges in the image spaces of multiple images.Some points, especially those grid points in non-texture or littletexture areas, will fail in MPGC matching. These points are alsokept but they are assigned low reliability indicator values.

4. EXPERIMENTAL RESULTSIn order to evaluate the performance of our approach for DSMgeneration it has been applied to different areas with varyingtextures, terrain types and image scales. In the following we willreport about DSM accuracy test results with SI, IKONOS andSPOT5 HRS images.

4.1 SI Image Dataset, GSI area, JapanIn Japan's GSI (Geographical Survey Institute) test area, both SIimages and aerial photos are available. The evaluation is basedon the comparison between the manually measured DSM fromaerial photos and the automatically extracted DSM from the SIimages.

The GSI test area is roughly 650 × 2500 m2. It is relatively flatwith natural and artificial objects. There are a lot of smallgeomorphological features, but also small discontinuities likecars, isolated trees and large discontinuities and occlusions dueto buildings. For the success of a matcher it is very importanthow it can handle local discontinuities (e.g. buildings or otherman-made objects, vertical cliffs, etc.). Figure 5 shows two image windows from the nadir-viewing SIimages. We have 48 control points. They are signalized markson the ground or on the top of buildings. They appear both in theSI and aerial photos.Two stereo pairs of color aerial images of 1:8000 image scale,acquired with a film camera of 153 mm focal length, have beenused for manual collection of reference data on an AnalyticalPlotter. The RMSE of the exterior orientation is reported as 5 cmin planimetry and 3 cm in height. Manually measured points,distributed at about 50 cm distance in both X and Y directions,are used as reference data.Assuming a measurement accuracy of 0.1% of the flying height(best case situation for natural points) we can expect an accuracyof the reference data of 0.12 m as a best case scenario.

Figure 5: Region (left) of man-made objects. Region (right) of bareterrain area (including some sparse trees). The image patches are

from the nadir SI-100 images with the ground resolution of 5.6 cm.

Figure 6: 3D visualization of the shaded DSM of the GSI test area.

Table 1: DSM accuracy evaluation.Manually measurement minus automatically extracted DSM

The processed SI data include three panchromatic images with afootprint of ca. 5.6 cm. As a result of triangulation, 2.8 cm and5.0 cm absolute accuracy in planimetry and height (as computedfrom checkpoints) were obtained.For analysis of the matching accuracy, we divided the referencedata into two classes, i.e. (1) the bare terrain area with somesparse trees and small artificial objects (including the largeparking areas); (2) the area with man-made objects and trees.Our analysis has been performed for these two classesseparately.A very dense raster DSM with 15 cm interval was interpolatedfrom the automatically matched point cloud and edges. Thepoints with low reliability values were given a small weight inthe interpolation procedure. Figure 6 shows the 3D visualizationof the extracted DSM.Comparing the reference points with the raster DSM in twodifferent areas (1) and (2) leads to the results shown in Table 1.The bare terrain area shows much better results, but it stillsuffers from problems like multi-temporal differences between

aerial photos and SI images, and some sparse trees. In the man-made object area, the accuracy becomes worse. Here 2.4 % of allpoints points have more than 4 meters difference. These pointsare almost all positioned on the border of buildings and trees.This may be caused by the fact that manually measured pointsand automatically extracted points are referring to differentobject points due to small errors in orientation procedure and 3Dmodeling problems.

4.2 IKONOS Image Dataset, Thun, SwitzerlandThe test covers the area around the town of Thun, Switzerland. Itis about 17 × 20 km2. The terrain elevations range from 600 m to2200 m, with some very steep and high cliffs.One IKONOS stereo pair was acquired on 2003-12-11 10:29GMT over an area of approximately 11 × 20 km2. AnotherIKONOS triplet was acquired on 2003-12-25, 10:39 GMT overarea of approximately 11 × 20 km2. The two areas have nearly50 % overlap. All images are IKONOS Geo products. Thesensor and sun elevation angles (ca. 19 degrees) were less thanoptimal. The low elevation angle of the sun causes strongshadows, especially in the south part of the images and ingeneral low contrast images. A 2 m resolution reference DSMgenerated from airborne LIDAR in the year 2000 was obtainedfrom the Swiss Federal Office of Topography, Bern. Thereference DSM only covers the south part of the whole area.

Figure 7: Extracted DSMTop: The whole area; Bottom: Town of Thun

Firstly, the original 12 bit images were preprocessed and resultedin enhanced images for further image matching. Then the imageswere orientated with the help of about 40 GCPs measured byGPS. The orientation accuracy is about 0.41 m in planimetry and0.68 m in height. Finally, a 5 m raster DSM was generated fromthe matched mass points and the edges. Some areas like lakesand rivers are defined as “dead areas” manually.Figure 7 shows the 3D visualization of the generated DSM (thewhole test area and part of the town area). The results show thateven small geomorphological features are extracted and surfacediscontinuities are well preserved.

Table 2: DSM accuracy numbers (west part of test area).O-Open areas; C-City areas; V-Tree areas; A-Alpine areas.

Table 3: DSM accuracy numbers (east part of test area).O-Open areas; C-City areas; V-Tree areas; A-Alpine areas.

Table 2 and 3 give the DSM accuracy test results. We computethe differences between the interpolated heights from our DSMminus the heights of the reference DSM. The accuracy of thegenerated DSM is at 1.3 – 4.8 pixel level. It depends on theterrain type. Higher accuracy can be achieved in open areas. Inurban and tree areas the accuracy becomes worse. The analysisshows that points with more than 6 m differences are almost alldistributed in the tree and urban areas. In open areas (with somesparse trees and small cluster of houses), more than 70 percent ofthe points have differences of less than 1meter. The results showsignificant biases. This is caused by different point definitions inlaserscan and photogrammetric surface points. Also, the differentacquisition times of the IKONOS images and the laserscans mayplay a role . For more details see (Eisenbeiss, et al., 2004).

Table 4: Reference DEMs. Height accuracy in meter

Figure 8a: Color shaded extracted DSM (the whole test area)

Figure 8b: Reference DEM (left, 25 meter grid of dataset “DLR-DEM-05-2”) and the generated DSM (right, 25 meter grid)

4.3 SPOT5 HRS Image Dataset, Bavaria, GermanyHere we report the work carried out within the ISPRS–CNESInitiative on DEM generation from SPOT5-HRS stereo images.For details see (Poli et al., 2004)The test area (No. 9) covers a part of South Bavaria and a part ofAustria (approximately 120 × 60 km2). Table 4 givesinformation on the reference DEMs. A stereo pair from SPOT5-HRS was acquired on 1 October 2002 in the morning. The

images have the resolution of 5 m in along-track and 10 m inacross-track directions.The test area includes a mountainous area (rolling and stronglyinclined alpine area) and hilly areas (rough/smooth and weaklyinclined areas). Our image matching software not only generatesa large number of mass points, but also produces line features.The TIN based DSM was generated from the matched masspoints and the edges (as break-lines).Figure 8 shows the 3D visualization of the generated DSM. Theresults show that the shapes of our generated DSMs are similar tothe references, but slightly smoothed. This can be expectedbecause of the 5m resolution of the satellite images.Tables 5 and 6 show the DSM accuracy test results. Theorientation accuracy is about 6.3 m in planimetry and 2.6 m inheight. We compute the differences between the heights of thereference DEM and the interpolated heights from our DSM.Table 6 shows the DSM accuracy test result by masking out thetree areas manually.From Tables 5 and 6 it can be seen that:• The accuracy of the generated DSM is more or less at the1pixel level or even better. Only the datasets 5 give values atabout 2 pixels, but these higher values are mainly caused bytrees.• All datasets still contain some blunders, which our proceduresfailed to detect.• The results show systematic errors. In datasets 5-1 and 5-2 thebiases are about 1 pixel. Except in case of dataset 6 all biases aresignificantly negative. This indicates that our generated DSMsare higher than the reference DEMs, an effect which could beexpected..

Table 5: DSM accuracy, units are meter

Table 6. DSM accuracy, units are meter(excluding the tree covered areas)

5. CONCLUSIONSIn this paper we have reported about our current matchingapproaches for fully automated DSM generation from lineararray images with different resolutions. We have developed amatching strategy combining feature point matching, grid pointmatching with neighborhood smoothness constraints, and robustedge matching. The strategy allows us to bridge over areas withlittle or no texture and at the same time maintain the importantcontribution of object/image edges. The modified MPGC is usedto refine the matching results in order to achieve sub-pixelaccuracy. The geometrical constraints are derived from thespecific sensor models for the linear array imagery, which can bethe rigorous sensor model for aerial and satellite images or theRF (Rational Function) model for satellite images.As evidenced by a visual inspection of the results we canreproduce even small geomorphological features. The resultsfrom the quantitative accuracy test indicate that the presentedconcept leads to good results. If the bias introduced by trees andbuildings is taken out, we can expect a height accuracy of onepixel or even better from satellite imagery (e.g. IKONOS andSPOT) as “best case” scenario. In case of very high resolutionaerial images (footprint 8 cm and better) it is obvious that the“one pixel rule” cannot be maintained any more. Alone surface

roughness and modeling errors will lead to large deviations, suchthat an accuracy of three to five pixels should be considered agood result. This is at the same level as laser scanning results. Ofcourse, the photogrammetric data can also be produced with thesame or even better point density. On the other hand, with theseaccuracies we are still operating at a coarser level than withmanual measurements from analogue aerial images, but we dothat with the advantage of great gain in processing speed.A major problem left is the control and automated detection ofsmall blunders, which still infest the results, despite thesimultaneous matching of more than two images. Thisconstitutes a relevant topic for further research.

ACKNOWLEDGEMENTS. We appreciate the support ofthe Swiss Federal Office of Topography, Bern, which providedthe laserscan data. We also thank Henri Eisenbeiss, who helpedin setting up the Thun area as a testfield for highresolutionsatellite image processing.

REFERENCESBaltsavias, E. P., 1991, Multiphoto Geometrically Constrained

Matching. Dissertation, IGP, ETH Zürich, Mitteilungen No.49, 221 pages.

Eisenbeiss, H., Baltsavias, E., Pateraki, M., Zhang, L., 2004.Potential of IKONOS and QuickBird Imagery for PointPositioning, Orthoimage and DSM Generation. IAPRS, Vol.35, Istanbul, Turkey (to be published)

Gruen, A., 1985, Adaptive Least Squares Correlation: Apowerful Image Matching Technique. South Africa Journal ofPhotogrammetry, Remote Sensing and Cartography, 14 (3),pp. 175-187.

Gruen, A., Bär, S., Bührer, Th., 2000: DTMs DerivedAutomatically From DIPS - Where Do We Stand?Geoinformatics, Vol.3, No.5, July/August, pp. 36-39.

Gruen, A., Zhang, L., 2002. Sensor Modelling for Aerial MobileMapping with Three-Line-Scanner (TLS) Imagery. ISPRSCommission II Symposium on Integrated System for SpatialData Production, Xi'an, P. R. China, August 20 – 23.

Gruen, A., Zhang L., 2003. Automatic DTM Generation fromTLS data. Optical 3-D Measurement Techniques VI, Vol. I,Zurich, pp. 93-105.

Kanade, T., Okutomi, M., 1994. A Stereo Matching Algorithmwith an Adaptive Window: Theory and Experiment. IEEETransactions on PAMI, Vol. 16, No. 9, pp. 920-932.

Murai, S., Matsumoto, Y., 2000. The Development of AirborneThree Line Scanner with High Accuracy INS and GPS forAnalysing Car Velocity Distribution. IAPRS, Vol. 33, PartB2, Amsterdam, pp. 416-421

Otto, G. P., Chau, T. K. W., 1988. A “Region-Growing”Algorithm for Matching of Terrain Images. Proc. 4th AlveyVision Club, University of Manchester, UK, 31 Aug. – 2Sept.

Poli, D., Zhang, L., Gruen, A., 2004. SPOT-5/HRS Stereo ImageOrientation and Automatic DSM Generation. IAPRS, Vol 35,B 1, Istanbul, Turkey (to be published)

Saint-Marc, P., Chen, J-S., Madioni, G., 1991. AdaptiveSmoothing: A General Tool for Early Vision. IEEETransactions on PAMI, Vol. 13, No. 6

Prazdny, K., 1985. Detection of binocular disparities. BiologicalCybernetics, Vol. 52, pp. 93-99.